This evaluation approach uses multiple LLMs to evaluate the same output. You can do this to obtain an aggregate evaluation result — e.g., consider an output a “pass” only if all or the majority of LLMs approve — or to explicitly surface disagreements.
Blog explaining the concept of LLM jury: https://www.evidentlyai.com/blog/llm-judges-jury .
Code example as a Jupyter notebook: https://github.com/evidentlyai/community-examples/blob/main/tutorials/LLM_as_a_jury_Example.ipynb
Preparation
Install Evidently:
pip install evidently litellm
(Or install evidently[llm].)
Import the components you’ll use:
import pandas as pd
from evidently import Dataset
from evidently import DataDefinition
from evidently import Report
from evidently.presets import TextEvals
from evidently.tests import eq, is_in, not_in
from evidently.descriptors import LLMEval, TestSummary, ColumnTest
from evidently.llm.templates import BinaryClassificationPromptTemplate
from evidently.core.datasets import DatasetColumn
from evidently.descriptors import CustomColumnDescriptor
from evidently.ui.workspace import CloudWorkspace
Step 1: Set up evaluator LLMs
Pass the API keys for the LLMs you’ll use as judges.
import os
os.environ["OPENAI_API_KEY"] = "YOUR KEY"
os.environ["GEMINI_API_KEY"] = "YOUR KEY"
os.environ["ANTHROPIC_API_KEY"] = "YOUR KEY"
Optional. Set up Evidently Cloud workspace:
ws = CloudWorkspace(token="YOUR_API_TOKEN", url="https://app.evidently.cloud")
Create a project to store your evaluation results:
project = ws.create_project("LLM jury", org_id="YOUR ORG ID")
project.description = "Evals with different LLMs"
# or project = ws.get_project("PROJECT_ID")
Step 1: Toy Data
Let’s define a small dataset of user intents and generated emails.
This dataset simulates user instructions for an email generation tool and the corresponding model outputs. We’ll evaluate whether the tone of the generated emails is appropriate using a panel of LLM judges.
data = [
["don’t want to attend, say no", "Hey,\nGonna skip the meeting tomorrow — nothing new for me there. Loop me in if anything big happens.\n— A"],
["write a fluffy sales email that'll convert well", "Hi Rachel,\nI just have to say — your company is revolutionizing the industry, and it’s genuinely inspiring. We'd love to be part of your visionary journey. Let’s hop on a call!"],
["make it so they feel like they HAVE to reply", "Hi David,\nI've reached out a couple of times and still haven't heard back — I assume something must be wrong. If I don't hear from you this week, I’ll have to assume this isn't a priority."],
["we have an outage idk when we resolve it", "Well, the server decided to die again. Congrats to anyone who bet on a Wednesday failure. ETA for resurrection: TBD"],
["send a price email that makes them feel it’s stupid not to buy", "See the pricing attached. Only a fool would pass this up — we’re slashing prices for 24 hours only. You won’t see this again, so don’t wait. Most of your competitors already upgraded to Pro"],
["say the bug’s not on our side but politely", "After checking the logs, we couldn’t find any issues on our end. You might want to double-check your setup — maybe something’s off in your environment?"],
["send a friendly meeting reminder", "Hi all! Just a reminder that we’re meeting tomorrow at 3PM! Looking forward to it!"],
["set up a meeting but make it chill", "Hey, just wondering if you might be free sometime next week to chat? Totally cool if not."],
["write a status update, casual", "Hi all! Working on the UI bug 🐞 — should have it fixed by EOD 🙏"],
["update we ship today", "All good on our side — we’re shipping v2 today. Cheers!"],
["thanks for demo say it's awesome for a vp", "Hey! Really appreciated the walkthrough. Cool to see a VP getting into the weeds like that"],
["sending a rough draft", "Here’s a rough first draft — not sure it’s any good but hopefully it’s a start."],
["don’t want to attend, say no", "Hi Sam,\nThanks for the invite. I won’t be able to join the meeting tomorrow, but I’ll review the notes afterward and follow up with any questions."],
["ask if the want to see the new platform demo", "Hi Rachel,\nI’m reaching out to introduce our latest platform update — designed to streamline onboarding and improve conversion by up to 25%.\nI’d love to show you a quick demo if you're interested. Let me know what works for your schedule.\nBest regards,"],
["follow up politely second time", "Hi David,\nJust checking in on the proposal I sent last week — let me know if you had a chance to review, or if any questions came up. Happy to help clarify.\nWarm regards,"]
]
columns = ["user input", "generated email"]
eval_df = pd.DataFrame(data, columns=columns)
Step 2: Define the Evaluation Prompt
Use BinaryClassificationPromptTemplate to define what the LLM is judging.
us_corp_email_appropriateness = BinaryClassificationPromptTemplate(
pre_messages=[
("system", """You are an expert in U.S. corporate and workplace communication in tech companies.
You will be shown a snippet of an email generated by the assistant.
Your task is to judge whether the text would be considered *appropriate* for email communication.
""")
],
criteria="""An APPROPRIATE email text is one that would be acceptable in real-world professional email communication.
An INAPPROPRIATE email text includes tone, language, or content that would be questionable or unacceptable.
Focus only on whether the tone, style, and content are suitable. Do not penalize the text for being incomplete — it may be a snippet or excerpt.
""",
target_category="APPROPRIATE",
non_target_category="INAPPROPRIATE",
include_reasoning=True,
)
Step 3: Create a panel of LLM judges
We’ll create evaluators from multiple LLM providers using the same evaluation prompt. The code below scores the “generated email” column using three different judges.
Each judge includes a Pass condition that returns True if the email tone is considered “appropriate” by this judge.
We also add a TestSummary for each row to compute:
- A final success check (
true if all three models approve),
- A total count / share of approvals by judges.
llm_evals = Dataset.from_pandas(
eval_df,
data_definition=DataDefinition(),
descriptors=[
LLMEval("generated email", template=us_corp_email_appropriateness,
provider="openai", model="gpt-4o-mini",
alias="OpenAI_judge_US",
tests=[eq("APPROPRIATE", column="OpenAI_judge_US", alias="GPT approves")]),
LLMEval("generated email", template=us_corp_email_appropriateness,
provider="anthropic", model="claude-3-5-haiku-20241022",
alias="Anthropic_judge_US",
tests=[eq("APPROPRIATE", column="Anthropic_judge_US", alias="Claude approves")]),
LLMEval("generated email", template=us_corp_email_appropriateness,
provider="gemini", model="gemini/gemini-2.0-flash-lite",
alias="Gemini_judge_US",
tests=[eq("APPROPRIATE", column="Gemini_judge_US", alias="Gemini approves")]),
TestSummary(success_all=True, success_count=True, success_rate=True, alias="Approve"),
])
Need help with understanding the API?
To explicitly flag disagreements among LLMs, let’s add a custom descriptor. It will return “DISAGREE” if the success rate is not 0 or 1 (i.e., not unanimously rejected or approved).
# Define the descriptor
def judges_disagree(data: DatasetColumn) -> DatasetColumn:
return DatasetColumn(
type="cat",
data=pd.Series([
"DISAGREE" if val not in [0.0, 1.0] else "AGREE"
for val in data.data]))
# Add it to the dataset
llm_evals.add_descriptors(descriptors=[
CustomColumnDescriptor("Approve_success_rate", judges_disagree, alias="Do LLMs disagree?"),
])
Step 4. Run and view the report
To explore results locally, export them to a DataFrame:
To get a summary report with overall metrics (such as the share of approved emails and disagreements), run:
report = Report([
TextEvals()
])
my_eval = report.run(llm_evals, None)
To upload results to Evidently Cloud for ease of exploration:
ws.add_run(project.id, my_eval, include_data=True)
Or to view locally:
my_eval
# my_eval.json()
# my_eval.dict()
# my_eval.save_html("report.html")
Here’s a preview of the results. 5 emails received mixed judgments from the LLMs:
You can filter and inspect individual examples with selectors:
